Scientific Methodology
The physics of molecular prediction. No shortcuts.
Every prediction in DrugSynq traces back to reproducible physics or ML models with documented validation datasets. We believe computational chemistry earns trust through transparent methodology, not marketing claims.
Free Energy Perturbation
Thermodynamic rigor for SAR navigation.
Binding free energy differences (ΔΔG) are computed by running alchemical transformations in thermodynamic ensembles. Unlike docking scores, FEP directly accounts for explicit solvent, protein flexibility, and entropic contributions to binding.
DrugSynq uses Hamiltonian replica exchange (HREMD) to enhance conformational sampling across intermediate states, reducing hysteresis in perturbation legs and converging results in fewer nanoseconds of aggregate simulation time.
Force Field
OPLS4: parameterized for drug-like scaffolds.
OPLS4 introduced new torsional parameters for heterocycles and heteroatom environments common in drug-like molecules, addressing systematic errors in earlier force fields for aromatic amines, sulfonamides, and fluorinated compounds.
Heterocycle Coverage
Extended torsional coverage for pyridines, imidazoles, pyrimidines, and other N-heterocycles that appear in 78% of FDA-approved small molecule drugs.
Halogen Bonding
Explicit sigma-hole parameterization for chlorine, bromine, and iodine enables accurate modeling of halogen bond-mediated binding interactions in kinase and protease targets.
Solvation Accuracy
Hydration free energies for a test set of 450 drug fragments show mean absolute error of 0.45 kcal/mol vs. experimental values — critical for absolute binding free energy baselines.
ML-Enhanced Potentials
Physics backbone, ML correction layer.
For targets where crystallographic data is available but the binding site exhibits conformational flexibility not captured by rigid-ligand docking, DrugSynq applies machine-learned correction terms trained on high-quality QM reference data.
This hybrid approach (OPLS4 + ΔML correction) improves accuracy by approximately 0.2–0.4 kcal/mol RMSE on prospective test sets for kinase targets relative to OPLS4 alone, without adding interpretability concerns of pure ML models.
See Published BenchmarksADMET Architecture
Ensemble models from in vitro training sets.
Each ADMET model is an ensemble of gradient-boosted tree classifiers trained on standardized in vitro assay data from curated public databases supplemented with internal measurements from collaborator labs.
Training Data Curation
All training data passes structure standardization, duplicate removal, activity cliff detection, and assay condition normalization before model training. 247,000 curated data points across 12 assay endpoints.
Prospective Validation
Model performance is evaluated on held-out prospective datasets (compounds not in training timeline) to assess generalization rather than interpolation. Reported metrics are prospective AUROC and MCC.
Quarterly Model Updates
ADMET models are retrained quarterly as new in vitro data is incorporated. Subscribers receive notification when model versions change, with side-by-side accuracy metrics for continuity.
Published Accuracy
Metrics you can reproduce.
Performance numbers are published with data splits, code, and test set SMILES. If you can't reproduce it, we consider the benchmark unreported.
| Endpoint | Prospective AUROC | Test MCC | n (test) |
|---|---|---|---|
| hERG Inhibition | 0.91 | 0.73 | 3,410 |
| CYP3A4 Inhibition | 0.88 | 0.68 | 5,820 |
| Metabolic Stability (HLM) | 0.85 | 0.61 | 2,190 |
| Aqueous Solubility | 0.87 | 0.66 | 4,750 |
| Caco-2 Permeability | 0.89 | 0.71 | 3,080 |
Published Research
Peer-reviewed evidence base.
Science you can build a drug program on.
Schedule a methodology review with our scientific team to evaluate fit for your target class and compound series.